AI Coding Live News

As autonomous agents execute complex multi-step workflows across terminal consoles and web browsers, context window expansion and escalating token costs impose severe operational bottlenecks. Conventional compaction methods rely on text summarization or token pruning, both of which discard crucial state variables and induce semantic drift over extended sessions. Researchers from HKUST and Alibaba present REMORY (arXiv:2610.11287), introducing a residual memory paradigm designed for sequence-level context compaction. REMORY combines lossy natural language summaries with bounded sets of learned soft memory tokens, effectively forming a residual connection in the sequence dimension. While text summaries convey high-level narrative continuity, the soft memory tokens reconstruct lost fine-grained execution artifacts, API arguments, and intermediate environment states. On challenging benchmarks including BrowseComp and Terminal-Bench 2.1, REMORY retains 97% of full-context task performance while occupying only 5.2% of the original input token footprint, slashing redundant tool retries and loops by more than 40%.

Designing optimal multi-agent harnesses—specifying roles, prompts, tool registries, and communication graphs—is foundational to task performance. However, because different user requests demand wildly disparate cognitive workflows, static hand-crafted multi-agent topologies fail to generalize; conversely, running exhaustive trial-and-error executions at inference time incurs prohibitive latency and cost. Researchers from Google Cloud AI Research, Georgia Tech, and UC Davis introduce SHIFT (arXiv:2610.04137), an architecture that decouples physical execution from the per-query search loop. SHIFT trains a local LLM architect policy over harness-building actions paired with a learned value function predicting the utility frontier between accuracy and execution token cost. For every incoming query, Monte Carlo Tree Search (MCTS) constructs a customized multi-agent topology tailored to the problem. Evaluated across 9,193 tasks spanning six diverse benchmarks with a Gemini 3.5 Flash executor, SHIFT achieves an unprecedented ~80% mean accuracy, beating 17 competitive baselines by up to 7.2 percentage points while slashing execution tokens by 32%.

@aiRobertDaily highlighted fast-jev-compaction, a Claude Code plugin that replaces traditional LLM summarization with Jev-driven, decision-based context compaction. It quickly filters tool calls and results, drops or truncates stale content, and keeps only the necessary parts of valuable text; a related Codex-facing Jev context server was also noted. The aim is lower token cost and faster delivery.

Guillaume Laforge (@glaforge) shipped Antigravity Brain Visualizer v0.6.0 for inspecting, understanding, and debugging Antigravity agent sessions. Highlights: single-pass Gemini 3.8 Flash session analysis, git checkpoint diffs, an artifacts explorer, and token/cost estimates.

Developer @midudev highlights open-source RTK (github.com/rtk-ai/rtk): it intercepts terminal command output and compresses it before it reaches coding agents such as Claude Code, Cursor, Copilot, and Codex, claiming about 70% token savings with automatic operation. Aimed at cutting token cost from verbose shell/log context in agent loops.

Open-source terminal tool Aider introduced Architect/Editor dual-model pairing. Complex reasoning and architectural decisions are handled by top-tier models (o3-mini / Claude 3.7 Sonnet), while diff editing and file patching are delegated to fast, economical models (DeepSeek-V3 / Qwen-Coder), slashing API costs by 45% while preserving 90%+ solve rates.

NVIDIA published the first on-silicon benchmark metrics for its Vera Rubin architecture on long-horizon agent workloads, demonstrating up to 30x throughput per megawatt and 35x lower token cost versus GB300 NVL72.

Google DeepMind publishes technical paper on Gemini 3.7 Flash Thinking. Benchmark results show the model reaches 68.0% on SWE-bench Verified across a native 1M context window at 1/20th the token cost of traditional frontier models.

DeepSeek previewed next-gen V4 architecture featuring sparse MoE, dropping long-context coding token cost to $0.07/1M.